kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-02-03 17:52:19 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[None][feat] update trtllm-gen to support groupsTokensHeadsQ (#10261)

Browse Source 
Signed-off-by: Perkz Zheng <67892460+PerkzZheng@users.noreply.github.com>
Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>
Co-authored-by: Bo Li <22713281+bobboli@users.noreply.github.com>
This commit is contained in:
Perkz Zheng 2026-01-15 15:24:25 +08:00 committed by GitHub
parent e12a7119cf
commit 71ccc07d2b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3836 changed files with 9419 additions and 14157 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
cpp/tensorrt_llm/common/attentionOp.cpp
View File

@ -1041,7 +1041,7 @@ int AttentionOp::mlaGeneration(
TllmGenFmhaRunnerParams tllmRunnerParams{};
// Parameters to select kernels.
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Dense;
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Causal;
tllmRunnerParams.mKernelType = FmhaKernelType::Generation;
tllmRunnerParams.mMultiCtasKvMode = mMultiBlockMode;
// Note that the tileScheduler and multiCtasKvMode will be automatically tuned when using multi_block mode.

4
cpp/tensorrt_llm/kernels/fmhaDispatcher.cpp
View File

@ -125,7 +125,7 @@ bool FmhaDispatcher::isSupported()
{
tllmRunnerParams.mSparseMla = true;
tllmRunnerParams.mKernelType = FmhaKernelType::Generation;
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Dense;
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Causal;
}
foundKernels = mTllmGenFMHARunner->isSupported(tllmRunnerParams);
@ -237,7 +237,7 @@ void FmhaDispatcher::run(MHARunnerParams runnerParams)
tllmRunnerParams.mSparseMla = true;
tllmRunnerParams.mSparseMlaTopK = runnerParams.sparse_params.sparse_mla_topk;
tllmRunnerParams.mKernelType = FmhaKernelType::Generation;
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Dense;
tllmRunnerParams.mMaskType = TrtllmGenAttentionMaskType::Causal;
tllmRunnerParams.kvPageIdxPtr
= reinterpret_cast<int const*>(runnerParams.sparse_params.sparse_attn_indices);
tllmRunnerParams.kvPtr = runnerParams.sparse_params.sparse_mla_kv_cache_pool;

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7e0e2b774a6c3cabb0e060ed69e2c8e2c6bbe1764bbbaab38d36e347e3bb57ce
size 609323
oid sha256:aab9ac4e7321815981e7f16588880c4471bc11d7ef847c5ba59823a61f7102c1
size 630932

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:83194aa60840a949f84496bb9f27659d4694d1e635935330d755523c36b735a8
size 545131
oid sha256:7c42a76a6d469e9981fd31517a0eea89ad25176eafd9ec847a90965e69e8da58
size 565951

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:458b8def069e59c69b6b9fdbc906eb3a245ea139229d5b2385c27a5ec5ae18c9
size 598271

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:97589b590015dc0349dc53d2349e76e5f747267310fab94600d0c4b6a22486cd
size 534081

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:62fadf416568ed8c5be310a53ac251a3488c00a2308b3cb0f9048d42e95e719b
size 460315
oid sha256:6d203f17a13605424fc336835b0da020a03b7822a2eb57f3b49180b26fc867c5
size 485079

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1efe0fce91ba26a5cad02bee190c24be2d6751b17d144c3f9a4af12931a3a39a
size 429011
oid sha256:26f4f24197d75d46fcbedfc9de1557a994800d47ce7b0c22cc04cb3927bc7706
size 454565

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1424bc3198804ac37471502af729d5ed30b4088c168bd45316342c3e2ae2d7ff
size 449579
oid sha256:d79548d5a0ab81b8bb4ab36aefda378bd908fff8be0bcd6b3c6363e29b5743dd
size 472765

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1cc5097176a862183eb1fa18d773bdfc245d9376744d28deebb7450ac7340e55
size 425403
oid sha256:344abd862e54d481b3c9edcd16218fef31f7a443fd7f2a9a839de743f6738145
size 450169

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a4b4d0561696d5728a80cd24af358c63ab78dc5b2feac10d43767236ae331547
size 605373
oid sha256:2885aa27b1f8efc168f3ebde84fc86678ab384e2e35ef5a3b208289e94690ed7
size 626982

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:801e9f9aaa027e98e7466db1e5a71ceea509e7faffdc5aeb9debb91b6db8f89a
size 543625
oid sha256:0270860fd490ab53e5a9fd28cbd4219ccaa4375ddd9919fe721dc1c6b3243a36
size 566023

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f5993b141d7cde0f588f496e8bb89df22d2ab6c359f3598077b9241747156ad4
size 428597
oid sha256:bfd79f0c8f79be828198380b0b1a0868d3168a447659dca8752cd9dd732f4503
size 452573

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:bc6584aed823757456b163a4140ea1660815760ff4f966fd0c36e32e06bff9a0
size 370623
oid sha256:bb30dc1e4c550621636e8c853e6b104723667fd2b48d5674c85f1c4617fa874f
size 394599

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:39d3d3da78f96164b8245e2ded87d2d5dc831ef5d1ece26cecb3c3aad178a79b
size 412315
oid sha256:e97c1dbc7556316ec7e35f078732736009c83e29fd7dffc0b9bb1449eb1da0f0
size 436291

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c70b84adee9466978f712c974654b53ca4d21eca3ba4ee7f3763a945679f4208
size 354341
oid sha256:6ed203011340282225246e72220500aa1d555f68036b786fc2db0ee0fe98806d
size 378317

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:056470956e537a019ab0019fc5aeb0b3e45814e8ac1c4f2545d79c2bacd5752f
size 449263

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6058dc20904d61781a8fdeb49a96cbaae4d4b0bdbc03a3b263886d19fa7d81c1
size 417959

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:25ee9eed542e3b6c6d09544b0a7fe77a4f3585ef9958b2c97dfd5e7b0ad5a6dc
size 438527

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d35c5308dbd338f71e815558e58438b59e6bacd1f9953e80eb82602b0b1c4830
size 414353

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d8abd45bd9ad702dbdaee33b3997f31616dd8af1a5bb847a7aeda34ef9d1a906
size 594323

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:134105ca624ddc23f161faec9c4b5401ef2f8246ed8f0816204cb8b5f97cd3b4
size 532575

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:da228e2856d9a716df7170b35702965f07bc9dff19196df782f399dc615991e8
size 418337

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4fc99a15e9a7b7db772277ae26457e391a3d5ff452eaca28f32ec21a61d1709c
size 359573

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:582516974c5b71a9794888940bbbe10f170c44e0f65a3f7d4bf1f7b048d12d59
size 401265

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:70d9d8daedd35ff863ce288284c0ebfd722665569f0b7f788d516e386720b169
size 343291

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9708057344f2813105b824dcdc494f5e05c9563d2a9545cb51a9aa0c5ab47761
size 479293
oid sha256:0c719151d9ec920ede687ea6e643bf6f8b4fc51d298c23dac9a9a480a9abfb4e
size 508005

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a268b9b5c2319938c604fabd9ec695905bae7432d9e00ed5af08c3d8fec4da31
size 448777
oid sha256:cf13e09558450edad962a3b2dcb40b241e19ae73fc094cb5e3b4d2e1aeb0f4ec
size 471175

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a008063d504f7beb4a6e275f40689da982deee6608778addab4ecd958903fa2f
size 469345
oid sha256:c94d9acd5903fc23f77a8f9a50725cd3702646d8e4cd1f8ac347bba2b71bd579
size 495689

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8acbba887bfb703aa796c0984143a281026db3720bbf6166cdd871f4fe9929c1
size 443591
oid sha256:d7afb5710156b1107f18c149ffc429770348d387c0976fdae13998d895ec5eac
size 471515

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c7f2258a99b512e91b896a897478e8fd9af91d09b26b335b1d515b75a8656ce3
size 647342
oid sha256:8141ce31fa80e0b34ee5103cc88dd081b9e46f9190a2c1417e0a76b852069bcc
size 670528

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:116e1a3c6610378ac7981f139e9bcdcbb5025d4df8c2b11a90949577aae2c08f
size 583595
oid sha256:d6d3c73fa9bab18a2fd5640d585e6456cdbf2cd2ce0c768c6dc96db7f4341565
size 607571

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:771eb850d08c939ce268c35a8992c40b8c233b66235533a19b22ca66c1c83762
size 452213
oid sha256:6fb4e40e49c327b4412be9d80d2fa36f69ee9b0f9f3a189cb69d231ab429cf52
size 481715

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:77982ca37706064117eb0bdc13d3deb99fa8ade3ca93f8778209633ed1dba118
size 383163
oid sha256:f5c464e1d6055c7b1b914ba28ffd46214dd067074f1ec5845c301dab56fef3b1
size 411875

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ab7b342275848b478e51eb310d670eb2ead7f0b6408a64fb7675736ed42ec997
size 432773
oid sha256:1283f9838b33699479b3cb220ae7a6daeaba289df8f6784da22226b65365c04c
size 460697

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:42f8da62dd959b5973cf4aea5da0d962381a023b80c3269e1a6981c7ef1deecb
size 366091
oid sha256:197830e1331734f431c27ccefbc6b8abd43b5cbdca9633ceff01de653409f37e
size 394015

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9720d9564adc6ff1d74d45573264964ec2bf8b2ecade6895c7ceb3c3749ab2ee
size 469031

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0b0a71145651cb709f558b450bcb2af2f845433bdbb6cd9c36e094744c8ae0dc
size 437727

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0c50e8ca669694c4adf0a2ef442e5ccf67be475ea65fec5942cc5adc102b1ac7
size 458295

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e0af13adfa5314867bed6f014dc96082f98563ca464ac8fe9dde77aecabe2e77
size 432541

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:fb125cf9a68c4843e35eec39071d589d86978b1fda5b54daaf757cb80fa93f85
size 636292

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ce48d46a620fb43941968c1b662f7b4a3a2e708d6ac8007bde57b4edf65220e3
size 572543

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:bacd646e2029c28a362aab6fae049fabdc04f6cf460bc914f54f1e29cef4d139
size 439583

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:63668716c020f026b8812907395cc82a3ff05ea89f258ba25cde1b136b30ae12
size 372113

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a20a7e2e15e62a8400d9f40d5fb0946cbbbf96906e5e42a7bc5b6b2d4c84b4ef
size 420145

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b1deae00bfd30a89b6df43a665d3f34934f221cc98e189e51d60033a4c9f9fe3
size 355041

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9c75d601fa7804e62bafb1dbb06c377a3d194a07d8e9eccd0276224023dd5c38
size 638998
oid sha256:44787d6712a551d25b917d99a6eb5fa13500ed03f2e26a279b341c88c4140902
size 663022

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a1eba3a874451f32716f7f283e7849e2b45fd89c049589453533515fc806ba1f
size 553617
oid sha256:07508bc40af84a09327533065d5a092465a5c2f206165c57a36f5149bb898804
size 574435

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:419cf33a68bcaf4bc1c6cad48d810be7ffd816ad0fd1b3eef94bcec0344bb5c9
size 617685

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:047335b382bd67c2262e5f8c058af1f7b3a7fc6a5e58bb72b8b962add25790e6
size 532305

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5e7b2dee7e6748c12b158c6223cd1b3e731d34b235456713b64e2e7defa68ffb
size 570821
oid sha256:f7a631d93550a5c4649669883cdc3bce68b0573ab5a6dad93f6748647ff42180
size 595587

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0f9b3a116151ea6aba238627d6760c5d306fea245ca77823935f2ac15b0f745f
size 545831
oid sha256:7165266f1f1f613f44ce22e05e4953348c3a75f32b7652ce14b21d6b76f0d3c0
size 571387

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:400db48260e6ff2bb4b6574195c63735c10a60e1e28861dc802caeb8bf24489b
size 557717
oid sha256:318e2c4385be1e2569c5702c56b66d9b8e223bb4bebd1e15eef9dbcc8c0f9029
size 581693

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:07bcaeae935460978795014e17b9b697dc6e2d1f653b866f2a53764bc51f570f
size 537489
oid sha256:3131c8d5f5fb015fcacd145c3017473825baefd64a60a624f3289030e05104f9
size 560675

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a5065c4697a089402900408bdc0e044fd615fe6bd0ac9cd4bd1677259f57bdb3
size 639588
oid sha256:024dc41efed33d2c38c11d18255009a9a3b2cc2800aee0d8aa81ad635448813f
size 663760

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e3ec5e0a5ba0f4f53e0cfb8aba7d33887655ba825d2df72598c8114f67e6baff
size 587679
oid sha256:f078613e55c6832616133eef2f68be51cca4e4a4a048edd2f8630c98763e9d64
size 608499

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:483979d9e9a46fe5c62e062aefe4e04f8820ebf3cb7b986f3249a94b4d0c08c6
size 534369
oid sha256:e018fa3ab404f4a2af235291ff638dcfbaee86f48a6eb2fff84b8642fd64c804
size 558345

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9a5716424ec1be6109841eaba73ea0e6a3bb1cf2882d78789426901157cd3a7b
size 477183
oid sha256:cbdbe20d955c73deff21abfa6107d8d49cc8560f5eb22d9ea9218ccd08ca3d5b
size 501159

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0084d18498fd64071bbde52dbd829c3735e3dd23ab73d11a69bd0793c2cf2065
size 515719
oid sha256:a14260c98eb9cbf4cd991370fb702da342a564a2090d0f2ba85917f20e801f11
size 539695

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:920dd8f1288e1005131021afab361245e7ff6bc5ebfc441ae0cfe92ed1df9b88
size 459323
oid sha256:4a92a1759f8ae7d760193a645c93ce7bfd54681a800839a0d8ddbc9ff8ec0ecd
size 483299

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ceff7493ab713c4e0a0ead3a6d1dd4c894ed9add166885c96950e855899045e4
size 550299

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7485cf9c19ab7de43f1af75924674cdf8c6af27ece1ca6e12c1e703dc77772d6
size 524519

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:73eee78418a2c922210e9063ca10bc3b51107fc5aacdd273320de6e301d970ba
size 536405

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:499b741bab50a4cae71db27a6a275c3ffe8708716d8ad3cceb811d67c69b0973
size 516177

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c3d5cc8682a111e81fa688af081d70f268d9798cc5d7cd2036098af8c402a0f7
size 618276

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f8bae3b4682e75004584192ed991be710abc5fcac2f0eb7166ce3283ec430e59
size 567157

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2c8c3fb8697b5f82be3d3617f2b281b76db12209c0a19719a7d0ddc619c58b34
size 513057

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e2af58295e357cd025c2d41d916f94d78ff4b9c2241f80b1d5c6e6dd9f427764
size 455871

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0c3c0492a243211464328106ff758dcd8f49b70b1788175b83ba2553fa32040f
size 494407

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c7be04907db7df6890b8ff47477d92537e97a060b2a15e716f74b961a50ca293
size 438799

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:207ce5a7b77fc1369a8065e32726a5907f831ef716f3a07c1a57f7efa1e74f69
size 590589
oid sha256:39712104e975e87c0a6b4a7702a9adebce0c13d429271ed20c0e461c57f0535b
size 618512

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0c95ea63c1fd0cf9be1f2d4a0bd00cbcaef293947e8bb2e649c8831730fbfe6c
size 564809
oid sha256:c133abb655b3c4951333b48115f8641eeb4d380245c20565fcd5c0ef55c5c4ff
size 593521

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5335c20af9df1e135acadda4e8500fc93a9a6fd57f242ba786eaec29a9362ed1
size 578273
oid sha256:fad53d9ba155ce085e9933b5b5dfba6fc07e153d65cb200a26c3a4a5bc6199bf
size 604617

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:61482aa0efa3b1e2bc5de30371004e8e3dce895b79e64d031773d8d9c3f331c4
size 555677
oid sha256:aa632239be24a5a2707fc6f6e85bc83479072e1294cd9eac2c8fb47993c6bd16
size 582811

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:89a7be0324a89c1c012aa5bde5fc27b09102440f13625e1c8cdf4c8de2b8e28c
size 677016
oid sha256:ea57bea5a2fc471860e4c5525fdcae29470dda4fe1300f99036585586d35a140
size 702670

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:36ed9309120c2bb74d26fc261da0647f85cd7c578aa9e2f909ecac16e31d5cd2
size 592425
oid sha256:c91939f73ae90335e7233a1b8bd732ed3cc94d178172388bab60bb7d01aeed9b
size 617191

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4215710fc336a28d5e4a8c862f4a8c65219704faad0e45fcd52496a1f931aaf4
size 557195
oid sha256:b46ecbb25482181e34dbe16854abe1c5a688608834821c5679c057c5fb7c43d3
size 586695

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:97b6587efd2ca2da444fc664f8b588146b8683b86d41f59ff836592a5fe9aea9
size 484987
oid sha256:1bd636e50b2cae9c7cfb5da70a1f59545cf5efef7ecd00f9036f6dce48e2b98b
size 512909

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:06ebf1bdaf037345c1ea881c9708f1caa2e9f9122cf277b173e6f45019a21a7d
size 536177
oid sha256:a1894d55f3eca9f0d9299fb9704a07ed69e3dfc9c35cc6e2e5b2da2dc35c6e97
size 564889

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d0cbd364d7223f22f63e8b14d059018af229dd00f646edaf4a83ce753622c24d
size 466337
oid sha256:c98aec7849f9db765d3edfb979f219aa5419466b9e1e54eea67ec9f5df58c0bc
size 494259

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:78a48e288b5ebf69af44e0d198d712515162a775a7bfa0bde9778ee4984c6003
size 570065

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:455753d48f62a8c3f09ec80f12d9d6e31c6b51f5a33a249ff285a68148f7f9a0
size 544287

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:55cc933af0d609b5f444b63bd260177165e37b424a4ed693be443e0ccca78666
size 556961

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:958fc38e924aa8ea21a1df82e28ad2af31b258a34105771736e61fed00e72ba9
size 535155

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f4570ab05595afe4f4a5024f4c9d1ae2b0a89a17612a7c76189607625915fa3f
size 655704

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:599119a25d21588f2efeccfedc79938505e399d6d05af78c5da0437e5d27b675
size 571903

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2f3d00ac5deb0c52d6485c2743956c2c4d32c625da4920f74ad12c2a25955dd2
size 534303

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a19cf3c9047cfcbaa0272699859342c4d4cb2999245858d79154fa103a30f711
size 463675

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3c87395de8198a6c252acdba4e8feb285a467c4a721f983ea1cd0018112644c6
size 513285

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cfa1a5b253a8ce83cbaf5f1d4a89c7c82029dd99e7b43f85a37bef9ec34fb06a
size 445815

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:850d27cac2525410314f6e9095dff09d88c80e0e5658adb4ccb1f04a45fcf4b6
size 540673
oid sha256:83c25ec85687ea926e665fbeb370dec55d36412238318013b0a392f87fb577fc
size 561491

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c57c7e6c626460af41f0ee2e20ae5e7f79e3e8b4309751fb0c7b557fe3d77bf6
size 482131
oid sha256:75804fb6424217fcaf15ce78a87c0629ecc35b6a3a1f43eddad1354691c581af
size 502949

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2af9cd8eeda9e5f233d5927e2693d62dbec67b0127db77934816f0b7646759ff
size 529623

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3a21ced2993682127666a369714bfc6d07663d093fadd682522bbafb0d23143f
size 471081

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3c75ef38ea0423414ff5739740c42694aa2f36c82eb59b7d1f89b64f0a0b8ddb
size 433697
oid sha256:6e86ec832ea89bd957499fcbf2d38d86b8dc3170e0e657cd94d83fec34f0a155
size 459251

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5bb999a86972033e51335c25bc5e6dda1e4512d21446645bff8552644386f3bd
size 418747
oid sha256:be19e95ebd0d58474840670c8d616b56eeebde025e3b4853bd59f206ec5666a9
size 444301

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:dfbe2d3f7dbd2e18f0ec5df77582edcf585a6c3844821276b203bfd5a4a58cfa
size 428461
oid sha256:524a093590794c4cf8f42cf6e6fa639bfd5e6d86be5050f0f538341f9937ef5d
size 453227

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1732ea77c95feef233dba9189c850db55059bcc446031449a082098c78d1cb59
size 413561
oid sha256:b90becb8d6b7abc262309e319f7fa6a6acc46946be96ced0219f1a762544776d
size 438327

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7a7943be47f715e598f5ddb45e6dd4fe2f2391475d949297d7734877205eee6d
size 535145
oid sha256:bc0433873f87e978d8ee8262d23a8726e469935fbeb3266861343b42f4cc9f4f
size 558333

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:08c46f3f1873b47bf8ed31ced47ed044d0293acaaed88dc5619ad3f92ccb9967
size 480625
oid sha256:4619457c5cca7dfec5595fe210944782362dcb2967cef79367dde603f11dbe8a
size 502233

Some files were not shown because too many files have changed in this diff Show More